The way in which British AI policy has been developed in recent years strikes at the heart of a core debate about where power lies in Whitehall.
During the Starmer administration, and especially the Sunak era, the prevailing view amongst those shaping AI policy was that what most MPs and ministers thought about AI wasn’t what drove government.
The key players that mattered were the PM and bits of No. 10, the Cabinet Secretary, a few key Secretaries of State, and a small number of highly effective officials who largely worked with special advisers and the private office. A small, politically empowered centre can achieve outsized policy change by assembling unusually capable officials and giving them direct access, authority and protection from the wider system.
This approach fell squarely within the Cummings theory of Whitehall. Indeed, many of the people involved in working on AI policy at the time were in a similar orbit to the misfits and weirdos. And it has largely proven highly effective in recent years, while the majority of other aspects of British industrial policy have struggled.
MPs and most ministers, frankly, did not care about AI in those early years. They had not reckoned with the scale of impact that the technology was going to have: gradually, then suddenly. What interest there was didn’t register within Whitehall very much.
The early Bletchley days are an excellent example of this theory of power in action. There were lots of aspects of the Summit work that were hugely contested within the wider tech policy ecosystem. Only around 100 orgs were invited to the Summit, to reflect the specific focus on frontier risks from bio, cyber, and loss of control: the major risks that we are seeing warning shots from today. But a lot of civil society felt spurned that they weren’t at the table. They were frustrated that other risks were not being covered. I recall a well-respected civil society leader talking at the 2023 Labour Party Conference that year, saying that we should scrap the £100m spending on AISI and put it all into ‘funding AI ethics’. That was before you got to the debate about Britain (rightly) inviting China to Bletchley.
Each of these decisions was highly contested at the time. But it didn’t matter in the end. The Bletchley agenda did not sway very much. And that was because the ‘Cummings theory of Whitehall’ had proven correct. Sunak was aligned. Senior leaders from the Foundation Model Taskforce (which became AISI) reported to him every couple of weeks. There were some extremely effective spads who got the machine working and had a very strong sense of direction.
Largely, this theory of change was sustained in the Starmer era as well. As the demands of the AI response grew, more of Whitehall needed to be taken on a journey. But again, it was the officials, not the ministers or MPs, who were at the wheel.
An evolution of the Starmer era from Sunak was the hollowing out of the No. 10 operation. There was now a great No. 10 AI team in place, but the overall weakness of the Starmer No. 10 operation meant that they could not adjudicate between competing departmental demands.
This became increasingly important for AI, where issues such as building AI data centres, copyright, and public service adoption required not just DSIT, and saw departments adopt very different positions on what to do. Where there were particularly competent Secretaries of State, such as at DESNZ, this actually scuppered the AI agenda in the absence of a strong PM operation. But by and large, UK AI flourished under Starmer, thanks to the compounding effect of building exceptional state capacity targeted at a clear theory of change, enabled by the authority of a handful of ministers.
This theory of power, that it rests with the officials and not the politicians, is now about to be tested immensely.
All MPs are now expected to have a response on AI. The level of political concern demands it.
With hindsight, I think the people working in UK AI policy who largely subscribe to the Cummings view have failed to sufficiently engage a broader suite of MPs on AI. I count myself in this group. And I think it is could be coming back to bite these folks.
At the same time, other campaign groups, ones that I consider to have net negative policy positions, have been putting in the yards engaging backbench MPs and the House of Lords. This has played out already in a few ways, with Lords championing the Kidron amendment on copyright and the latest Superintelligence Bill shaped by groups such as ControlAI. Both bits of prospective legislation would, in my opinion, be damaging, and in the case of Kidron, have already helped to weaken our ability to train frontier models in this country.
We are now in an era where parliamentary scrutiny of AI is mainstream, and entrepreneurial backbench MPs can have a real impact. Many MPs will be looking at the way in which Bernie Sanders is shaping the national conversation in America and wondering if they could have a similar scale of impact. While the quality of parliamentary debate on AI has largely been woeful, there are some very thoughtful MPs who I hope will play a role steering the country in the right direction.
Largely, the first phase of UK AI policy rewarded elite executive coordination. A major question is the extent to which AI politics now increasingly rewards coalition-building across the executive, Parliament and organised outside groups.
I still believe that the Cummings theory of power will endure, especially as the securitisation of AI continues on current trends. The strength of No. 10 and the quality of officials working on AI will determine, in my view, where the direction of UK AI policy is decided. But it is clear that the theory of change needs to update, in my view, towards new kinds of coalition building.